20VC: OpenAI and Anthropic Threatened by Kimi? | Should the US Ban Chinese Open-Source Models | Should Openrouter Sell & Value in the Routing Layer? | Stripe Buying Paypal: What You Need to Know

Summary

Harry Stebbings and Jason Lamkin debate how Chinese open-weight releases like Kimi and Qwen are pressuring frontier AI pricing, but they caution that benchmark wins are not the same as real deployment performance. They argue that regulated customers and agentic workflows can actually increase token consumption, yet the economics of training, inference, and rapid model obsolescence keep pushing vendors toward lower prices and heavier infrastructure spend. OpenRouter is framed as a likely seller because model-routing is becoming embedded and commoditized across vendor stacks, while Fireworks is presented as a beneficiary of the open-weight wave with strong revenue growth, high token throughput, and plans to move deeper into the data-center layer. Later, the conversation broadens to Stripe’s possible acquisition of PayPal, public-vs-private staging for companies like Databricks and nuclear startups, late-stage venture valuation compression, and how AI-driven demand is reverberating through semiconductors and memory pricing.

Notes

• Chinese open-weight releases are treated as near-frontier, but the speakers insist benchmarks alone do not prove field performance.
• They frame Kimi and similar models as technically impressive, but still distinct from smaller runnable-on-a-PC open models.
• Harry says Chinese model releases are still about six to nine months behind Western frontier models, depending on measurement.
• The episode treats the rise of cheaper equivalent models as a durable market force, not a one-off.
• Harry argues the open-weight opportunity implies roughly 80% cheaper intelligence once inference costs are separated from IP.
• The speakers discuss whether the U.S. should restrict Chinese open-source models, with pricing and industrial policy colliding.
• Jason says U.S. penetration by China-built models will be limited, but the size of that limit is still unclear.
• The conversation links regulated workflows to heavier agentic use because supervisors and parallel agents reduce error.
• Jason says that supervision stacks can multiply token usage rather than reduce it.
• Harry points to a gap in U.S. execution, arguing several domestic companies could build comparable open models but are not exploiting the opportunity.
• OpenRouter is described as being in talks to be acquired by multiple buyers.
• Jason explicitly says OpenRouter is a rational seller because the routing layer is getting embedded into many vendors.
• He argues the routing layer is becoming a default feature rather than a standalone product.
• Harry sees strategic value in model-agnostic routing if it helps a hyperscaler shift enterprise share.
• The speakers repeatedly argue that the infrastructure layer is where the best capital allocation is showing up.
• Harry’s mental model divides AI into infrastructure, foundation models, and apps, with infrastructure dominating spend.
• Jason says the application layer still does not feel like the real center of AI value creation.
• Custom labeling and domain-specific workflows can dramatically improve outputs versus generic model use.
• Jason says generic models are merely a starting point, while bespoke labeling produces much better workflow-specific systems.
• He describes the generic LLM as a blend of mediocre experts, while the best experts create a step-function improvement.
• Harry asks whether customer-built models add to or cannibalize foundation-model growth.
• He says the answer to that question would move markets, especially if it changes the trajectory of OpenAI and Anthropic.
• Jason argues competition is forcing model providers to respond, including through access changes and pricing pressure.
• Harry emphasizes that AI vendors face hard physical economics, unlike old software with essentially zero COGS.
• Jason says the core issue for model vendors is pricing, even if that means building chips and the whole stack.
• Consumer subscriptions are framed as heavily bundled and subsidized compared with enterprise token use.
• The Stripe-PayPal segment begins from the view that the transaction is strategically obvious if price clears.
• Harry frames the bid as a scale play: Stripe is much larger in valuation and could double its footprint through PayPal.
• Jason says the headline synergy is real, but the growth-rate math matters more.
• Harry says private-company financing complicates such a deal because stock issuance is limited.
• He says Delaware-style board duties require a good-faith belief that the company can do better than the offer on a standalone basis.
• Jason says the initial 28% premium looks pre-scripted because take-private deals usually move toward the mid-30s.
• Harry says Stripe buying PayPal feels inevitable from a long-run strategic perspective.
• Harry argues the broader AI conversation is dominated by OpenAI and Anthropic because they represent a huge share of private-market value.
• Harry notes that AI is also pulling demand through adjacent semiconductor leaders like TSMC and ASML.
• The pair describe late-stage AI investing as a market where pricing remains high, but multiples are still compressing.
• Jason says Series B investors now need something closer to seed-stage picking skill.
• Harry says later-stage investing benefits from both lower dispersion and the post-2022 equity bid.
• Jason says he would be uneasy as a founder doing contemporaneously tranched rounds.
• Jason says he would feel uncomfortable charging different investors dramatically different prices in the same week.
• Harry says tranche pricing is mostly about bragging rights and deterring rival VCs.
• He adds that two-tranche rounds are imperfect, but less dangerous than heavy structure or debt.
• Harry argues that nuclear startups should probably remain private, while Stripe and Databricks should probably be public.
• Jason describes C Square as a slow-growth AI-adjacent data-center business with a $1B run rate and $3B market cap.
• Jason’s takeaway is that the market is not stupid; companies still need to deliver revenue growth.
• Harry contrasts long-term semiconductor cooperation with the brutally competitive DRAM market.
• He says DRAM suppliers are behaving like a commodity cartel, with 40% quarterly price hikes and very high margins.
• Harry says rising memory prices are lifting AI infrastructure costs by roughly 2x.
• Nvidia’s valuation is framed around future AI capex rather than the past ChatGPT step-up.
• Harry says the most important variable is OpenAI and Anthropic growth in 2026 and 2027.
• Harry says the wider AI market is still highly concentrated in a few frontier-model companies.
• Jason says venture pricing at hot AI companies still shows real multiple compression.
• Harry’s ad read claims Base44 can build apps, websites, and AI agents from plain language with backend, database, auth, and hosting included.
• The Plaud read says the device is card-sized and already has more than 2 million users.
• The Fin read says the product uses custom models trained on years of customer interactions.